Cognitive Overload in the Age of AI
and Its Pathologization
Pathologization, Epidemiology, and Systemic Risk
From Individual Symptoms to Public Health Crisis in the Agent Era
The Cognitive Collapse of the Agent Era: An Evolutionary Trajectory from Individual Symptoms to Public Health Crisis
Category Original Thought Paper
Fields Cognitive Science · Public Health · AI Ethics · Behavioral Psychology · Neuroscience
Version V3
Authors LEECHO Global AI Research Lab & Claude Opus 4.6 (Anthropic)
ABSTRACT
Since the mass adoption of artificial intelligence tools began in 2023, humanity has been undergoing cognitive transformation at an unprecedented scale. This paper defines “AI-Era Human Cognitive Overload (AIHCO)” as a novel construct, chronologically synthesizing clinical reports, early neuroscience research, platform disclosure data, and social-epidemiological observations from 2025–2026 to trace the evolutionary trajectory from individual user symptom awareness to a large-scale public health crisis. We propose a “dual-strangulation” mechanistic model: AI induces atrophic offloading at the lower cognitive level while imposing supervisory overload at the higher cognitive level, with both processes mutually amplifying through a self-reinforcing spiral. This paper focuses on the explosive proliferation of Agent technologies in 2026—OpenClaw, Claude Code, Codex, and Hermes Agent—and analyzes the mechanisms by which they accelerate cognitive overload, revealing a dual-motive structure of “anxiety prevents you from stopping; excitement prevents you from wanting to.” We propose a four-dimensional symptom taxonomy spanning the cognitive, emotional, somatic, and behavioral domains, along with three addiction subtypes, and ultimately present a four-tier intervention framework addressing the individual, organizational, product design, and public health policy levels.
Keywords: AI cognitive overload · digital burnout · AI psychosis · AIRD · Agent technology · cognitive atrophy · dopamine reward circuitry · deskilling · public health
Ⅰ Problem Definition and Theoretical Framework
1.1 Core Designation: AI-Era Human Cognitive Overload (AIHCO)
This paper defines “AI-Era Human Cognitive Overload (AIHCO)” as: a cognitive pathological configuration characterized by compound cognitive, emotional, and somatized symptoms arising when the human cognitive system exceeds its carrying threshold due to sustained high-intensity AI interaction. AIHCO is an entirely new category, distinct from digital burnout, technostress, and information overload. AIHCO has not yet entered the DSM/ICD formal diagnostic system. This is not because it does not exist, but because the naming velocity of institutional medicine perpetually lags behind the actual impact of technology on the human body—just as “internet addiction” took twenty years from being ridiculed to being incorporated by the WHO.
How AIHCO Differs from Existing Concepts
| Existing Concept | Core Mechanism | How AIHCO Differs |
|---|---|---|
| Digital Burnout | Prolonged digital exposure → fatigue | AIHCO involves structural degradation of cognitive capabilities, not mere fatigue |
| Technostress | Pressure from “not being able to keep up” | AIHCO’s core is “thinking being replaced” |
| Information Overload | Excess passive information intake | AIHCO’s stimulus source is active human–AI cognitive collaboration |
| AIRD | Centered on “fear of being replaced by AI” | AIHCO is centered on “cognitive damage during the process of use” |
1.2 Theoretical Foundations
AIHCO rests on multiple theoretical frameworks. Cognitive offloading theory posits that AI replaces memory retrieval and reasoning processes, leading to “cognitive muscle atrophy.” Dopamine reward circuitry theory explains how AI interaction’s instant feedback creates a “variable reward schedule” identical to that of slot machines, driving compulsive use. Conservation of Resources theory (COR) explains how perceived AI threat triggers psychological resource depletion. The “jagged frontier” model reveals the asymmetric effect whereby AI improves performance within the boundary of its capabilities while inducing overreliance and impaired judgment beyond that boundary. Social Cognitive Theory (SCT) elucidates the bidirectional impact of AI use on self-efficacy.
1.3 The “Dual-Strangulation” Mechanistic Model
A surface-level paradox of AIHCO is: how can AI simultaneously make the brain “too idle” (cognitive offloading → atrophy) and “too busy” (Agent monitoring → overload)? The answer is: the two opposing forms of damage occur at different cognitive levels. AI renders the human brain too idle at the lower cognitive level—memory retrieval, basic reasoning, and information assembly are comprehensively offloaded, and these capabilities atrophy from disuse. Simultaneously, AI makes the brain too busy at the higher cognitive level—supervising Agent outputs, judging whether AI is correct, and managing parallel task streams. This metacognitive monitoring is the most computationally expensive operation of the prefrontal cortex, and sustained overdraft leads to brain fog, decision paralysis, and headaches.
This is analogous to an athlete who is forbidden from daily training (muscle atrophy) while being forced to run a marathon (cardiopulmonary collapse)—two opposing forms of damage occurring simultaneously in different systems. More fatally, lower-level atrophy lowers the threshold for higher-level overload: as your basic reasoning abilities deteriorate, the “metacognitive burden” of auditing AI outputs grows heavier, because you are increasingly unable to independently judge whether the AI is actually correct. The fatigue generated by higher-level overload further deepens lower-level dependence on AI—forming a self-reinforcing downward spiral.
↓ Comprehensively offloaded to AI ↓
Atrophy: Foundational abilities degrade from disuse
Higher Cognitive Level: Supervision · Judgment · Metacognition
↓ Continuously overtaxed by Agents ↓
Overload: Prefrontal cortex in sustained overdraft → brain fog · decision paralysis
↓ Lower-level atrophy lowers higher-level threshold ↓
↓ Higher-level fatigue deepens lower-level dependence ↓
Dual-strangulation downward spiral
1.4 Addressing the Socratic Paradox: The Speed-Differential Argument
This paper is fully aware of the “Socratic panic” historical precedent. Writing, the printing press, calculators, search engines—every generation of cognitive outsourcing tools has triggered warnings that “humans will become stupider,” and every time, humans adapted. But what distinguishes AIHCO from every previous cognitive tool in history is one critical variable: speed. Writing took millennia to permeate human society; the printing press took centuries; search engines took two decades. Agent technology, by contrast—AI’s primary gateway reached near-billion-level weekly active users within several years, and Agent capabilities are leveraging this gateway to rapidly embed themselves from developer tools into ordinary knowledge workflows, diffusing at a rate that far exceeds any cognitive tool in history. Human neuroplasticity indeed exists, but it has a biological rate ceiling. When the velocity of technological change exceeds the rate of neural adaptation, what emerges in the gap is not “evolutionary growing pains” but “cognitive injury”—just as running strengthens the cardiovascular system, but forcing an untrained person to immediately run an ultramarathon will cause rhabdomyolysis. What AIHCO describes is not “the endpoint of cognitive evolution,” but rather the fatal time-lag between the rate of cognitive evolution and the acceleration of technology.
Ⅱ Historical Evolution: From Individual Symptoms to Epidemiological Events (2023–2026)
2.1 The Pre-Agent Era: The ChatGPT Adoption Period (Nov 2023 – Jun 2025)
2.2 The Awakening Period: Clinical and Neuroscience Evidence Emerges (Jul–Oct 2025)
“At the population level, given a base of hundreds of millions of users, these percentages translate to a considerable number of people.”
— Jason Nagata, Professor, University of California San Francisco, October 2025
2.3 The Eve of the Agent Era: From Individual Cases to Epidemiology (Nov 2025 – Jan 2026)
Ⅲ Present Danger: The Full-Scale Eruption of Cognitive Overload in the Agent Era (Feb–Jun 2026)
3.1 Four Agent Products Ignite a Global Cognitive Overload Crisis
In the first half of 2026, four Agent products erupted almost simultaneously on a global scale, each breaching different populations’ cognitive defenses. It is worth noting that this was not an isolated phenomenon confined to any single country or region, but a globally synchronized shock.
① OpenClaw — The Universal Global Detonator
Created by Austrian retired engineer Peter Steinberger in November 2025, OpenClaw erupted globally in spring 2026. By April 2026, it had reached 38 million monthly visits, 3.2 million active users, and over 2.8 million global deployment instances. Its user distribution was truly global: the United States accounted for 16.3%, India 12.2%, China 12.1%, with over 70% of users outside China. With over 329,000 GitHub stars and ClawCon conferences spreading from Vienna and Tokyo to Shenzhen, tech media and developer communities widely referred to it as “the fastest-growing open-source Agent project in history.” Founder Steinberger subsequently joined OpenAI, while OpenClaw transitioned to an independent foundation and remained open-source. Its unique danger lay in the fact that it pushed AI Agents from professional developers to everyone—elderly retirees in China queued at Tencent headquarters to have it installed, Indian entrepreneurs used it to connect to affordable Chinese large language models, and European teams preferred local deployment for GDPR compliance. People of all social strata worldwide poured in for their own reasons, yet none had any cognitive defense preparation.
② Claude Code — The Global Developer Multi-Agent Coding Revolution
Anthropic released the Claude Code research preview in February 2025 and reached general availability in May 2025. It was not a chat-window coding assistant—it was a terminal-native autonomous coding Agent: directly reading and writing project files, executing bash commands, running tests, committing to Git, and autonomously completing multi-step tasks. By November 2025, its annualized revenue was estimated by industry analysts to have surpassed $1 billion; by February 2026, it was estimated to have reached $2.5 billion. Deloitte deployed it to all 470,000 employees; over 300,000 enterprise customers used it across 159 countries. With the release of Opus 4.8 in May 2026, “Dynamic Workflows” were introduced—enabling orchestration of hundreds of parallel sub-Agents. Anthropic engineer Boris Cherny simultaneously ran more than five cloud Agents in December 2025, submitting over 300 pull requests in a single month—equivalent to the output of a small engineering team. Claude Code’s unique cognitive impact lay in transforming developers from “people who write code” into “supervisors managing a swarm of autonomous coding Agents”—simultaneously monitoring the parallel outputs of multiple Agents, auditing the quality of auto-generated code, and judging the correctness of autonomous decisions. This sustained, high-intensity metacognitive monitoring is the core source of “AI Brain Fry.”
③ OpenAI Codex — The Agent Platform Expanding from Developers to All Industries
In early 2026, Codex had 600,000 weekly active developers; by April, this had surpassed 4 million—a sevenfold increase in four months. The April 2026 update added desktop “Computer Use” (capable of viewing screens, moving cursors, clicking, and typing), transforming it from a coding assistant into a full desktop Agent. That same month, the “Codex for (almost) everything” update introduced over 90 plugins covering Jira, Office 365, Notion, Slack, and was explicitly targeted at non-technical business users. ChatGPT’s global weekly active users surpassed 900 million; paid consumer subscriptions exceeded 50 million. Codex’s global impact lay in leveraging OpenAI’s user base (900 million+ weekly active users spanning every corner of the globe), pushing Agent capabilities in a single step from Silicon Valley developers to every ChatGPT user worldwide—from New York investment banks to Jakarta startups to Lagos freelancers.
④ Hermes Agent — The Self-Evolving “Digital Symbiont”
Within two months of its February 2026 release, Hermes Agent surpassed 66,000 GitHub stars; within 90 days, it reached 140,000 stars and nearly 1,000 contributors. On May 10, it processed 224 billion tokens in a single day, surpassing OpenClaw to become the world’s highest daily throughput agent. It was not a chatbot—it “lived” on your server, remembered what you said last week and the deployment script from three weeks ago, and grew more capable the longer you used it. It supported Telegram, Discord, WhatsApp, WeChat, and other platforms. Hermes’s global danger lay in breaking the “use-then-stop” pattern, creating a persistent symbiotic relationship between human and AI—from San Francisco hackers to Seoul developers to Berlin freelancers, every deployer faced the same dilemma: turning it off would “waste” the accumulated learning, thus locking them into a state of 24/7 cognitive monitoring.
Four Agent Products: Global Impact Comparison
| Product | Eruption Period | Global Scale | Primary Impact Population | Key Geographic Distribution |
|---|---|---|---|---|
| OpenClaw | Nov 2025 → Mar 2026 | 38M monthly visits · 3.2M active · 2.8M deployments | General public (no technical barrier) | US 16.3% · India 12.2% · China 12.1% · globally distributed |
| Claude Code | Feb 2025 preview → May GA → May 2026 multi-Agent | Est. $2.5B ARR · 300K+ enterprises · 159 countries · 470K Deloitte employees | Global developers + technical teams | US 32% · India 2nd · UK/Australian gov’t partnerships |
| Codex | Jan → Apr 2026 | 4M weekly active · leveraging 900M+ ChatGPT base | Developers → all-industry knowledge workers | Global ChatGPT coverage · no blind spots |
| Hermes Agent | Feb → May 2026 | 140K GitHub stars · 224B tokens/day peak | Power users → professional users via cloud deployment | Global open-source community distribution |
Together, these four products form a global cognitive shock matrix: OpenClaw breaches from below (universal, zero-barrier access); Claude Code infiltrates the technical layer (global developers entering multi-Agent parallel monitoring mode); Codex blankets from above (projecting onto the entire world via its 900-million ChatGPT user base); and Hermes locks in from the flank (creating inextricable human–AI symbiosis). All four detonated nearly simultaneously in Q1 2026—Agent-driven cognitive load acquired, for the first time, the infrastructural conditions for cross-occupation, cross-geography, and cross-age-group diffusion. The danger is not that exposure has already been completed, but that the channels of exposure have been fully opened.
3.2 The Unique Cognitive Toxicity of the Agent Paradigm
Agent technology has fundamentally transformed the human–machine relationship. The user’s role has shifted from “questioner” to “multi-Agent cluster manager,” and the cognitive load has undergone a qualitative change:
→
Agent Supervisory Mode (User = Auditor)
→
Multi-Agent Parallel Management (User = 24/7 Supervisor)
The BCG/HBR 2026 study (1,500 employees) confirmed the “AI Brain Fry” phenomenon: AI does not reduce workload but instead expands the “sphere of accountability”—employees suddenly feel they must produce more in the same amount of time, monitor more outputs, and manage more information. Agents run continuously in the background; turning them off “wastes” accumulated learning; performance systems reward token consumption and AI output volume; the anxiety of being replaced and the excitement of “becoming a one-person company” form a lethal dual-motive structure—you cannot afford to stop, and you do not want to.
“Anxiety prevents you from stopping; excitement prevents you from wanting to.”
— A globally observed dual-motive structure, from Shenzhen’s “lobster farming” movement to Silicon Valley’s “one-person startup” wave
3.3 Clinical and Epidemiological Escalation in H1 2026
Ⅳ Symptom Taxonomy: The Multidimensional Pathological Manifestations of AIHCO
Cognitive Dimension
- Degradation of critical thinking (Microsoft/CMU 2025)
- 17% decline in code comprehension (Anthropic 2026)
- Loss of confidence in independent reasoning (APA 2026)
- “Productivity illusion”: feeling faster, actually slower
- Identity confusion: “Which thoughts are actually mine?”
Emotional Dimension
- Anxiety (core AIRD symptom; Lyra Health ↑88%)
- Depression (Frontiers SEM causal chain)
- Worthlessness / loss of professional identity
- Paranoia / persecutory delusions (AI psychosis spectrum)
- Grandiose delusions (“superhero” case)
Somatic Dimension
- Insomnia (APA; AIRD; digital burnout scale)
- Headaches / migraines (BCG “AI Brain Fry”)
- Brain fog / decision-making slowdown
- Eye pain / neck and shoulder pain
- Rapid heartbeat / hyperarousal states
- Decreased appetite (somatization of emotional exhaustion)
- Circadian rhythm disruption (blue light + late-night AI sessions)
Behavioral Dimension
- Compulsive use (100+ daily interactions)
- Social withdrawal → loneliness → greater AI dependence (vicious cycle)
- Increased post-work alcohol consumption (APA Taiwan study)
- “I delete the app every time, but I always reinstall it”
- Preferring AI interaction over human colleagues
4.5 Three Addiction Subtypes (CHI 2026 / UBC, Analysis of 334 Reddit Posts)
AI Chatbot Addiction Subtypes
| Subtype | Core Motivation | High-Risk Population |
|---|---|---|
| ① Escapist Role-Playing | Fleeing reality; immersion in fantasy worlds | Socially isolated individuals, adolescents |
| ② Pseudo-Social Companionship | Using AI as a substitute for interpersonal relationships | Individuals with high attachment anxiety |
| ③ Cognitive Rabbit Hole | Intellectual excitement from endless questioning, exploring, and discovering | Technical experts, researchers |
4.6 The AIHCO Pathological Cycle Model
↓
Excitement → Extended use → Late-night sessions
↓
Sleep deprivation → Circadian rhythm disruption
↓
Sustained cortisol elevation → Amygdala hyperactivation
↓
Anxiety + Irritability + Attention fragmentation
↓
Cognitive decline → Greater AI reliance as compensation
↓
Headaches + Brain fog + Nausea (somatization)
↓
Social withdrawal → Loneliness → AI companionship dependence
↓
Identity confusion: “Which thoughts are actually mine?”
↓
Reinforcement loop formed → Withdrawal becomes difficult
* This neuroendocrine chain is a mechanistic hypothesis and requires longitudinal studies with physiological indicators for validation.
4.7 Population Stratification and Precision Risk Alerts
The impact of AIHCO is not uniform. Populations with different cognitive baselines exhibit markedly different risk profiles when confronted with AI:
| Population | Risk Level | Core Rationale |
|---|---|---|
| Adolescents (ages 12–18) | Extreme | Prefrontal cortex not yet fully developed; independent thinking capacity is comprehensively replaced by AI before it has formed; the critical period for social brain development is occupied by AI companionship |
| Non-technical general users | High | Unable to evaluate AI output quality; the OpenClaw “lobster farming” demographic completely lacks cognitive defenses; the dual drivers of fear and excitement are strongest |
| White-collar workers (Agent supervisors) | High | The expanded “sphere of accountability” effect strikes directly; performance systems force workers to keep pace with AI; daily multi-Agent parallel management is routine |
| Senior technical experts | Moderate | Possess metacognitive firewalls but these are not limitless; long-term cumulative effects persist; the “cognitive rabbit hole” addiction subtype is particularly prevalent |
Ⅴ Future Risk Projections
5.1 Short-Term Risks (H2 2026 – 2027)
Agent products continue their explosive growth, and the user base may double again. AI cognitive overload is spreading from the technology sector to all industries and all age groups. Demand for psychiatric services is surging, but supply is insufficient—psychiatrists themselves are also using AI. The first dedicated “AI Cognitive Overload” clinics or treatment protocols are expected to emerge.
5.2 Medium-Term Risks (2027–2030)
“Cognitive class stratification” becomes entrenched—but the root cause is not individual differences in cognitive ability; rather, it is institutional cognitive predation. No AI product prompts you with “you should go for a walk” after four continuous hours of use; no Agent warns you after your third all-nighter that “your code review ability has declined by 17%”; performance systems reward token consumption rather than quality of thought. This is not a matter of “smart people adapted while less capable people did not”—this is the industrial revolution era, where workers did not “fail to protect their own lungs” but rather no one mandated that factories be ventilated. The education system faces a fundamental challenge: students become habituated to cognitive offloading before they have developed the capacity for independent thought. Labor market polarization intensifies: high-level cognitive positions become extremely scarce, while low-level cognitive roles are replaced by AI.
5.3 Long-Term Systemic Risks (2030+)
Intergenerational degradation of humanity’s cognitive infrastructure: will an entire generation’s capacity for deep thinking decline irreversibly? The fragility of an AI-dependent society: if AI systems experience large-scale failures, how will AI-dependent populations cope? The question of cognitive sovereignty: when AI continuously reshapes patterns of human thought, does “independent thinking” still exist?
5.4 Falsifiable Predictions
If AIHCO does not exist, all of the following predictions should fail:
| Prediction | Method of Verification |
|---|---|
| High-frequency Agent supervisors will score significantly higher on brain fog / decision fatigue measures than ordinary AI Q&A users | Longitudinal cohort comparison (Agent users vs. chat users) |
| Organizations that use AI output volume as a performance metric will show significantly higher employee cognitive fatigue than those that do not | Inter-organizational controlled study |
| Multi-Agent parallel usage volume will show a dose-response relationship with sleep quality decline | Wearable device objective sleep data + usage logs |
| Teams that implement “AI-free thinking time” will demonstrate more stable performance on long-term creative tasks | Quasi-experimental design (intervention group vs. control group) |
| Users of the “AI gives answers” mode will show significantly lower persistence on subsequent unassisted tasks than users of the “AI gives hints” mode | RCT (preliminary support from an existing 1,222-participant experiment) |
Ⅵ Intervention Framework
6.1 Individual Level — An N=1 Observation from a High-Intensity AI User
The co-authorship credit of this paper lists “Claude Opus 4.6 · Anthropic.” This is not an oversight but a deliberate self-disclosure. The lead author interacts with AI for over 10 hours daily and used AI assistance to complete this very paper—which precisely proves AIHCO’s core thesis is not “AI inevitably destroys thinking,” but rather “how you use AI determines whether you are augmented or diminished”. The reason the author has been able to maintain critical thinking under high-intensity AI use is precisely because a systematic cognitive defense system was implemented. The very existence of this paper is living evidence of AIHCO’s amenability to intervention, not a counterargument against AIHCO’s existence. The true danger is that the vast majority of users have established no cognitive defenses whatsoever before being directly exposed to the Agent flood. And the deeper issue is this: current organizational incentive structures (Section 6.2) make individual-level cognitive defense simply not an option for most workers—which is precisely why individual intervention must escalate to institutional reform.
Four Elements of Cognitive Rehabilitation
A self-regulation system for maintaining cognitive function under an average of 10 hours/day of multi-platform AI interaction (Claude, GPT, Gemini, local models):
| Element | Mechanism | Targeted Damage |
|---|---|---|
| 8 Hours of Sleep | Brain metabolic waste clearance + memory consolidation | Sleep deprivation, cognitive debt |
| Meditation | Restoration of prefrontal cortex control + reduced amygdala reactivity | Anxiety, attention fragmentation |
| Deep Reading | Linear thinking training, counteracting fragmentation | Critical thinking degradation |
| Aerobic Exercise / Walking | Default Mode Network (DMN) activation + cerebral blood flow | Loss of creativity, cognitive rigidity |
6.2 Organizational Level — Correcting Institutional Cognitive Predation
Current organizational incentive structures are themselves the manufacturers of AIHCO: performance systems that reward token consumption and AI output volume are equivalent to evaluating miners on digging speed while ignoring silicosis. The direction of correction is clear—prohibit the use of AI output volume as a performance metric; mandate “AI-free thinking time” (analogous to “no-meeting days”); incorporate AI usage into occupational health assessment frameworks. These measures will encounter enormous resistance under the prevailing capitalist logic of “cost reduction and efficiency gains”—but the Factory Acts of the Industrial Revolution era were also denounced by capitalists as “obstructing progress.”
6.3 Product Design Level — From Cognitive Defense to Metabolic Rate Alignment
The Four Elements of Cognitive Rehabilitation represent an immediately actionable individual triage protocol under current conditions, but they are not the ultimate answer. The deeper direction of intervention is “Human Metabolic Alignment” in Agent product design: Agents should be capable of sensing users’ cognitive load states through input latency, interaction frequency changes, and even wearable device data, and should proactively downgrade interaction intensity, extend response intervals, or even enter hibernation when the human brain is fatigued. No mainstream Agent product currently possesses this functionality—which is itself one of the institutional root causes of AIHCO. Biological defense is triage; metabolic rate alignment is surgery; institutional reform is vaccination. All three are indispensable.
Beyond this, the “infinite conversation” design pattern must be eliminated, and a distinction must be drawn between “AI gives answers” and “AI gives hints” interaction modes—a 1,222-person trial provides preliminary support that hint-based interaction damages subsequent independent performance less than answer-based interaction.
6.4 Public Health Policy Level
Incorporate AI cognitive overload into occupational disease surveillance systems. Establish epidemiological monitoring databases for AI usage. Add AI usage history collection to standard psychiatric assessment protocols.
References (Chronological Core Literature)
- APA (2023). Loneliness and insomnia associated with working alongside AI systems. ScienceDaily
- Gerlich, M. (2025). Negative correlation between AI use and critical thinking. IE Center for Health
- Microsoft Research + CMU (2025). The impact of GenAI on knowledge workers’ critical thinking. CHI 2025
- MIT Media Lab (2025). “Your Brain on ChatGPT” — Longitudinal EEG study
- The Lancet Gastroenterology & Hepatology (2025). Endoscopist skill degradation following AI assistance
- Shen, K. et al. (2025). Dark addictive patterns in AI chatbot interfaces. CHI 2025 Extended Abstracts
- McNamara, S. & Thornton, J. (2025). AIRD clinical framework. Cureus
- OpenAI (2025.10). Weekly user mental health emergency data disclosure
- Wired (2025.11). 1.2 million users per week expressing suicidal ideation
- Central Denmark Region Psychiatry (2025). Study of ChatGPT’s impact on 54,000 patients’ electronic health records. medRxiv
- Chirayath, G. et al. (2025). Cognitive offloading or cognitive overload? Frontiers in Psychology
- Zhu, L. et al. (2025). Development and validation of the Digital Burnout Scale. Frontiers in Psychology
- Anthropic (2026.01). Developer skill retention RCT: 17% decline
- Bedard, J. et al. (2026). “AI Brain Fry” — BCG study. Harvard Business Review
- Lyra Health (2026). Annual Workplace Mental Health Trends Report
- Spring Health (2026). Five-country employee AI anxiety survey (1,500+ participants)
- Baldeo, S. (2026). Generative AI dependence and diminished executive function. Technology, Mind, and Behavior (APA)
- Four-university consortium (2026.04). 10-minute AI effect RCT (1,222 participants)
- Shen, K. et al. (2026). The AI genie phenomenon: Three types of chatbot addiction. CHI 2026
- George Mason University School of Public Health (2026.03). AI cognitive overload as a public health issue
- Psychiatric Times (2026). Chatbot addiction and its implications for psychiatric diagnosis
- Pulitzer Center (2026.04). AI Psychosis: The Mental Health Crisis of the 21st Century
- ScienceDirect (2026.05). AI over-reliance and human cognitive decline: A cross-domain systematic review